Machine learning models with time-series clinical features to predict radiographic progression in patients with ankylosing spondylitis

doi:10.21203/rs.3.rs-934182/v2

Download PDF

Research Article

Machine learning models with time-series clinical features to predict radiographic progression in patients with ankylosing spondylitis

https://doi.org/10.21203/rs.3.rs-934182/v2

This work is licensed under a CC BY 4.0 License

Journal Publication

published 20 Dec, 2023

Read the published version in Journal of Rheumatic Diseases →

Version 2

posted

You are reading this latest preprint version

Background

Ankylosing spondylitis is chronic inflammatory arthritis that causes structural damage to the spine due to repeated and continuous inflammation over a long period of time. The purpose of this study was to establish the application of machine learning models for predicting radiographic progression in patients with AS using time-series data from electronic medical records (EMRs).

Methods

EMR data, including baseline characteristics, laboratory finding, drug administration, and modified Stoke Ankylosing Spondylitis Spine Score (mSASSS), were collected from 1,123 AS patients who were followed up for 18 years at a common center at the time of first (T₁), second (T₂), and third (T₃) visits. The radiographic progression of the (n + 1)^th visit (P_n+1 = (mSASSS_n+1 – mSASSS_n) / (T_n+1 – T_n) ≥ 1 unit per year) was predicted using follow-up visit datasets from T₁ to T_n. Three machine learning methods (logistic regression with least absolute shrinkage and selection operation, random forest, and extreme gradient boosting algorithms) with three-fold cross-validation were used.

Results

The random forest model using the T₁ EMR dataset showed the highest performance in predicting the radioactive progression P₂ among all the machine learning models tested. The mean accuracy and the area under the curves were 73.73% and 0.79, respectively. Among the variables of T₁, the most important variables for predicting radiographic progression were in the order of total mSASSS, age, and alkaline phosphatase.

Conclusion

Prognosis predictive models using time-series data showed reasonable performance with clinical features of the first visit dataset for predicting radiographic progression. Additional feature data such as spine radiographs or life-log data may improve the performance of these models.

ankylosing spondylitis

machine learning

radiographic progression

Ankylosing spondylitis (AS) is a chronic inflammatory arthritis that primarily causes ankylosis of the spine [1]. Patients with AS have chronic inflammatory back pain and gradually develop ankylosis of the spine, limiting their movement. Because structural changes due to inflammation may impact normal functioning and quality of life, the acceleration of radiographic progression is important for patients with AS.

In previous studies, statistical methods were generally used to investigate patient features related to spinal structural changes shown by radiography. They identified that radiographic progression was significantly correlated with men, tobacco, inflammation, and HLB-27 [2–4]. However, it is difficult to predict radiographic progression in an individual patient based on those factors because it depends on various other factors over a long period. Considering that numerous data with various types have been accumulated in the form of electronic medical records (EMRs) for AS patients under clinical care over time, statistical methods may have limitations in the analysis and prediction of radiographic progression in AS. However, machine learning methods can help predict radiographic progression using data accumulated over a long period and facilitate understanding of the complex relationships between variables in big data.

The use of machine learning methods in relation to big data has been on the increase in the medical field. The results of predictions made by analyzing data using machine learning methods can provide us with insights into the relationships between variables. Therefore, EMRs stored over time may be the best source for applying machine learning models [5]. Although machine learning models have recently been proposed to predict radiographic progression as well as disease activity, treatment response, and diagnosis in AS [6–13], the models have shown different performances owing to differences in the type and amount of data, hyperparameter tuning, and outcome settings.

A major strength of machine learning models is that they can handle complex and heterogeneous data such as time-series EMRs. This study aimed to explore the application of machine learning models to predict radiographic progression in patients with AS based on time-series data for each visit and to identify predictive datasets and important features that contribute to radiographic progression in those machine learning models.

Patients

This paper describes a retrospective longitudinal study conducted at a single center. The dataset comprised reviewed EMR data of 1,280 patients between January 2001 and December 2018. All patients were previously diagnosed with AS according to the modified New York criteria [14]. The Hanyang University Seoul Hospital institutional review board approved this study (HYUH 2020-03-012-003). Informed consent was waived because this study retrospectively reviewed the EMRs. This study included only anonymized patient data and all research was performed in accordance with the Declaration of Helsinki.

Clinical data

Clinical characteristics including age, sex, disease duration from the first to the last follow-up, HLA-B27 positivity, eye involvement with uveitis, and peripheral joint involvement with arthritis other than axial joints were investigated. Baseline laboratory results consisted of hemoglobin (Hb), hematocrit (Hct), blood urea nitrogen (BUN), creatinine, aspartate transaminase (AST), alanine transaminase (ALT), alkaline phosphatase (ALP), albumin, cholesterol, protein, creatine phosphokinase (CPK), gamma glutamyl peptidase (GGT), lactate dehydrogenase (LDH), erythrocyte sedimentation rate (ESR), and C-reactive protein (CRP) levels. The prescribed drugs were classified as follows: non-steroidal anti-inflammatory drugs (NSAIDs), methotrexate, steroids, sulfasalazine, and biological disease-modifying antirheumatic drugs (bDMARDs). The mean values of laboratory tests, the sum value of prescribed medication, and clinical characteristics were used as machine learning features.

Assessment of radiographic progression

Two radiologists (SL and KBJ) independently assessed the images and scored them according to the mSASSS (0–72) [15]. Intraobserver reliability with consistency for a reader was excellent (intraclass coefficient (ICC) = 0.978, 95% CI 0.976 to 0.979) and interobserver reliability with agreement between two readers was also excellent (ICC = 0.946, 95% CI 0.941 to 0.950) [16, 17].

Model design

The first, second, and third visits (T₁, T₂, and T₃, respectively) were defined as the time points at which the first, second, and third radiographs were taken, respectively. In addition, the radiographic progression at each time point was calculated using the following formula: P_n+1 = (mSASSS _n+1 − mSASSS _n) / (T_n+1 − T_n). radiographic progressor was defined as worsening of the mean mSASSS by more than one unit over 1 year, and AS patients were categorized into two groups: progressor and non-progressor [18].

We composed three clinical datasets for the prediction of radiographic progression: baseline dataset at first visit (T₁) with radiographic progression at second visit (P₂), two-points dataset at first and second visits (T₁ + T₂) with radiographic progression at third visit (P₃), and three-points dataset at first, second, and third visit (T₁ + T₂ + T₃) with radiographic progression at fourth visit (P₄). Using three clinical dataset matrixes, we trained three prediction models for progressor and non-progressor groups (Fig. 1). Three different machine learning classifiers were applied: logistic regression with least absolute shrinkage and selection operation (LASSO) using Python with the Scikit-learn package (https://github.com/scikit-learn/scikit-learn) [19], random forest (RF) using the Scikit-learn package [20], and extreme gradient boosting (XGBoost) using the Xgboost package (https://github.com/dmlc/xgboost) [21]. The algorithms were selected based on their superior performance and readiness for application. All continuous clinical features were centered and scaled to a mean of zero and a standard deviation of one (z-score transformation was performed before feature selection). The three models were run and compared to determine the best combination for determining progressor or non-progressor in the three clinical datasets. All possible combinations of the hyperparameters of the models were investigated by grid search using the GridSearchCV library in the Scikit-learn package [19].

With a linear combination of the selected features weighted by their respective coefficients, a LASSO regression model was used for prediction. RF, which is a representative ensemble method, is widely used because it is powerful and lighter than other ensemble methods. RF constructs several tree-type base models and forms an ensemble through a technique called bootstrap aggregating or bagging. XGBoost is a gradient boosted decision tree algorithm for large datasets. Detailed hyperparameters of the three models in the three datasets are described in the supplementary information (Supplementary Document).

Performance evaluation

The prediction models were evaluated in three rounds of three-fold cross-validation [22]. The procedures, including z-normalization and machine learning classification, were executed separately on the training data during each cross-validation. As the progressor and non-progressor groups are unequally distributed in the dataset, we used stratified cross-validation to divide the dataset. In each round, an entire dataset was randomly and equally divided into three with stratified probability. Two of these parts were used as the training dataset, and the final part was used as the test dataset. The process was repeated three times in three datasets of three models. The one-point dataset for predicting radiographic progression at second visit (T₁ for P₂) had 29 features, two-point dataset for radiographic progression at third visit (T₁ + T₂ for P₃) had 53 features, and three-point dataset for radiographic progression at fourth visit (T₁ + T₂ + T₃ for P₄) had 77 features. Each average of three models of the three-fold cross-validation in the one-point dataset became the estimated performance of the models. The same was the case for the other two datasets. The predictive power of each predictor was assessed through receiver-operator characteristics (ROC).

Feature selection

Features with greater contributions to the LASSO regression model were selected. We performed feature importance analysis using RF and XGBoost to verify the robustness of the results. Variable importance was evaluated using the model-based variable importance scores, and the important variables (particularly those informative to radiographic progression) were captured when the models were fitted to the training dataset [23, 24].

Statistical analysis

For continuous distributed data, the results are shown as means with standard deviations (SD); between-group comparisons were performed using Student’s t-test. Categorical or dichotomous variables were expressed as frequencies and percentages and were compared using the chi-squared test. Area under the curves (AUCs) were used to determine the diagnostic performance, with optimal thresholds of the clinical parameters determined by maximizing the sum of the sensitivity and 1 − specificity, i.e., the Youden index values. Machine learning model training and statistical analysis were performed using Python (Python Software Foundation, version 3.5.2)

Differences between non-progressor and progressor groups

Of the total of 1,280 patients, 157 patients did not have prescription and laboratory data; therefore, 1,123 patients were included in the study. The average time intervals between T₁ and T₂ and between T₂ and T₃ were 2.27 years and 2.12 years, respectively. The baseline characteristics of the non-progressor and progressor groups at the first visit (T₁) are shown in Table 1. The datasets of 1,123 patients at the first visit, 1,115 patients at the second visit, and 899 patients at the third visit were divided into training and test sets. (Fig. 2).

Table 1

Baseline characteristics in patients with non-progression and progression.
Variables	Total patients (n = 1123)	Non-progressor (n = 830)	Progressor (n = 293)	P
Male, n (%)	993 (88.42)	718 (86.51)	275 (93.86)	0.001
Age (years)	32.01 ± 9.41	30.98 ± 9.46	34.93 ± 8.65	< 0.001
Eye involvement, n (%)	363 (32.32)	245 (29.53)	118 (40.27)	< 0.001
Peripheral involvement, n (%)	401 (35.71)	319 (38.43)	82 (27.99)	0.002
HLA-B27, n (%)	1079 (96.08)	793 (95.54)	286 (97.61)	0.163
ALP	79.51 ± 32.98	77.82 ± 32.16	84.28 ± 34.82	0.005
ALT	21.55 ± 16.64	21.06 ± 16.91	22.95 ± 15.81	0.084
AST	19.96 ± 9.24	19.94 ± 9.36	20.00 ± 8.89	0.921
Albumin	4.33 ± 1.07	4.38 ± 1.01	4.19 ± 1.21	0.019
BUN	12.94 ± 4.69	13.14 ± 4.54	12.37 ± 5.06	0.022
CPK	96.40 ± 231.63	99.21 ± 243.98	88.43 ± 192.56	0.444
CRP	1.74 ± 2.09	1.61 ± 2.00	2.10 ± 2.31	0.001
Cholesterol	162.48 ± 50.70	162.31 ± 48.81	162.94 ± 55.79	0.864
Creatinine	0.83 ± 0.31	0.83 ± 0.22	0.83 ± 0.48	0.999
ESR	28.57 ± 27.30	26.89 ± 26.93	33.32 ± 27.82	< 0.001
GGT	14.97 ± 30.47	13.52 ± 26.00	19.09 ± 40.31	0.028
Hb	13.40 ± 3.17	13.46 ± 3.02	13.24 ± 3.57	0.346
Hct	40.64 ± 9.36	40.78 ± 8.88	40.23 ± 10.60	0.427
LDH	114.39 ± 77.68	115.65 ± 77.13	110.81 ± 79.23	0.366
NSAIDs, n (%)	880 (78.36)	650 (78.31)	230 (78.50)	0.987
bDMARDs, n (%)	246 (21.91)	185 (22.29)	61 (20.82)	0.659
Methotrexate, n (%)	151 (13.45)	122 (14.70)	29 (9.90)	0.049
steroids, n (%)	260 (23.15)	197 (23.73)	63 (21.50)	0.485
sulfasalazine, n (%)	283 (25.20)	228 (27.47)	55 (18.77)	0.004
mSASSS	14.57 ± 16.28	12.36 ± 16.07	20.84 ± 15.25	< 0.001
HLA, human leukocyte antigen; ALP, Alkaline phosphatase; AST, aspartate aminotransferase; ALT, alanine aminotransferase; BUN, Blood Urea Nitrogen; CPK, creatine phosphokinase; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; GGT, gamma glutamyl peptidase; Hb, hemoglobin; Hct, hematocrit; LDH, Lactate Dehydrogenase; NSAIDs, nonsteroidal anti-inflammatory drugs; bDMARDs, biologic disease modifying anti-rheumatic drugs; mSASSS, the modified Stoke Ankylosing Spondylitis Spinal Score.

Predicting radiographic progression with three time-point datasets

The radiographic progression was predicted using the clinical data at the first, second, and third visits (Table 2). Among the machine learning models, the RF model exhibited the best performance, with higher mean sensitivity, mean specificity, mean accuracy, and mean AUC than those of the lasso, logistic regression, and XGBoost models. In the RF model, the prediction of radiographic progression at second visit with first visit data (P₂ with T₁ dataset) showed better performance compared to prediction at third visit with first and second visit data (P₃ with T₁ + T₂ dataset) and at fourth visit with first, second, and third visit data (P₄ with T₁ + T₂ + T₃ dataset).

Table 2

Prediction performance evaluation according to time points and machine learning models.
Prediction of radiographic progression (Pn + 1) with visit data (Tn)	Lasso and logistic regression				Random forest				XGBoost
	Sensitivity (%)	Specificity (%)	Accuracy (%)	AUC	Sensitivity (%)	Specificity (%)	Accuracy (%)	AUC	Sensitivity (%)	Specificity (%)	Accuracy (%)	AUC
P2 with T1	68.25	68.31	68.3	0.7169	73.72	73.73	73.73	0.7959	70.99	70.84	70.88	0.7729
P3 with T1 + T2	66.18	66.3	66.27	0.6831	67.95	67.27	67.44	0.7467	66.21	66.3	66.28	0.7132
P4 with T1 + T2 + T3	61.39	60.03	60.4	0.6442	68.47	67.93	68.08	0.7348	66.8	67.94	67.63	0.7062
AUC, area under curve

The confusion matrix and ROC for the prediction of P₂ with T₁ dataset are shown in Fig. 3A and Fig. 3B, respectively. In three-fold cross-validation, the mean sensitivity, mean specificity, and mean accuracy were 73.72%, 73.73%, and 73.73%, respectively. The mean AUC of three-fold cross-validation was 0.7959. (Supplementary Figure S1 and S2 show the confusion matrix and ROC of LASSO regression and XGBoost model for P₂ with T₁ dataset, Figure S3A–S8A and Figure S3B–S8B show the confusion matrix and ROC of three machine learning models in P₃ with T₁ + T₂ dataset and P₄ with T₁ + T₂ + T₃ dataset).

Importance of features for predicting radiographic progression

The variables in the first visit data contributing to the prediction of radiographic progression at second visit using RF are listed in Fig. 3C. In three-fold cross-validation, the most important feature in common was total mSASSS, and the second and third most important features were age and ALP followed by CRP, cholesterol, ESR, hematocrit, and ALT. Drugs such as sulfasalazine and methotrexate, clinical features such as eye and peripheral involvement, sex, and HLA B27 contributed less to radiographic progression than laboratory findings. In the XGBoost model for P₂ with T₁, mSASSS was the most important feature in the variables; however, drugs such as sulfasalazine and methotrexate were also ranked high in feature importance (Supplementary Figure S2C). In addition, feature importance was identified in the RF and XGBoost models in P₃ with T₁ + T₂ (Supplementary Figure S4C and S5C) and P₄ with T₁ + T₂ + T₃ (Supplementary Figure S7C and S8C).

We developed a machine learning model that predicts radiographic progression using EMR data from patients with AS accumulated over 18 years. The RF model trained on data from the first visit predicted radiographic progression with an accuracy of 73.73% and an AUC of 0.7959, showing the best performance among the three machine learning models. We also found that the accuracy and AUC decreased as we trained with the second and third visit data. These results suggested that the data accumulated over a longer period did not increase the performance of the three machine learning models, and the data from the first visit in AS may contain important predictors for predicting radiographic progression.

Joo et al. predicted radiographic progression using machine learning on the training set (n = 253) and test set (n = 173) [9]. In their results, balanced accuracy in the test set was over 65% in all models and 69.3% in RF, which was the highest of all models. In addition, the generalized linear model and support vector machine showed the best performance with an AUC of over 0.78. The outcome of their study is similar to ours in predicting radiographic progression, but with significant differences in detail. First, we examined machine learning based prediction models for radiographic progression according to each visit using three time-point datasets containing EMR data accumulated over 18 years. We used a larger amount of time-series data compared to them and were able to identify clinical characteristics for each time point that could affect radiographic progression. These results provided various insights into radiographic progression in AS studies. In addition, the accuracy and AUC achieved in our study were higher than those achieved by them. This difference in predictive power may be related to the difference in the amount of data and variables, such as limited features for bone marrow density and syndesmophyte score and additional laboratory findings.

We used time-series EMR data accumulated from the first, second, and third visits to predict radiographic progression at subsequent visits. Interestingly, the predictive performance was poorer when the model included clinical datasets at the second or third visit compared to only the clinical dataset from the first visit. This could be explained by the fact that baseline data from the first visit contained important information for predicting radiographic progression. In addition, as treatment with NSAIDs was started at the first visit, the disease activity index such as the Bath Ankylosing Spondylitis Disease Activity Index, CRP, and ESR decreased subsequently. A decrease in the disease activity index, which leads to an increase in mSASSS [2–4], may have reduced the differences in important features between individuals. As a result, the prediction performance may have deteriorated for the datasets of the second and third visits.

The EMR data of AS have accumulated over time with various variables. EMR data are important in AS because the radiographic progression advances slowly over a long period and can be related to many variables. Above all, radiographic progression may be the result of the delayed effects of various clinical variables because inflammation begins, ossifies, progresses to syndesmophyte and can be confirmed on radiographs. Although disease activity, such as CRP or ankylosing spondylitis disease activity score, is known to be an important predictor of radiographic progression [25, 26], it is not an absolute long-term factor in determining radiographic progression. For example, radiographic progression continues even when recurrent transient inflammations are actively controlled [27]. This evidence suggests that various factors influence radiographic progression. Therefore, it might be difficult to find factors related to radiographic progression from the EMR data comprising numerous variables that we have using existing statistical analysis. Unlike the analysis of numerous statistical associations that have already been conducted, this study aims to provide insight into the timing and factors important for the prediction of radiographic progression.

Several machine learning models using large datasets have been useful for diagnosing axial spondyloarthritis [12]. Those approaches can help in early diagnosis and reducing the social burden of disease. Using a claim dataset, Deodhar et al. suggested that machine learning models have a positive predictive value of 6.24% compared to the Assessment of SpondyloArthritis International Society classification criteria with a positive predictive value of 1.29% [7]. In addition, machine learning models with EMR datasets have also shown good performance for predicting the diagnosis of axial spondyloarthritis with accuracies ranging from 82.6–91.8% [10, 11, 13]. It can be used for early diagnosis of AS by creating a machine learning model with image data and text data because images such as radiographs are important in the diagnosis of AS. The detection of sacroiliitis using X-ray, computed tomography, and magnetic resonance imaging using machine learning methods has been conducted recently with excellent performance in the screening of patients with AS [6, 8, 28]. Therefore, developing a machine learning model useful for diagnosis by combining image, life-log, and clinical information is essential to improve diagnosis accuracy, which is worthy of future challenges for the prediction of radiographic progression in patients with AS. Furthermore, an important task is assembling a representative and diverse dataset to meet the demands of high-performance machine learning models [29].

Machine learning methods may be useful even in areas with insufficient data [30]. Statistical methods employ a validation system that determines statistical significance based on the random sampling of data. By contrast, because machine learning models use big data, they rely on training data as a convenient sample rather than random sampling. Therefore, cross-validation or external validation is essential. Learning all the data in the population would solve this problem, but such an undertaking is infeasible. Nevertheless, machine learning models can be valuable because, once trained, a machine learning model can be easily changed to a model that reflects the characteristics of other hospitals with a small amount of data through “transfer learning.” However, caution is required in interpreting even the results inferred from these models.

There are some limitations to our study. First, we applied three machine learning models to predict individual radiographic progression and identified the importance of features that contribute to their prediction. Interpretation of the importance of features is possible because previous statistical studies have shown the factors related to radiographic progression. Therefore, machine learning methods may help complement statistical methods. However, additional statistical validation is needed for generalization of important unknown features that contribute to radiographic progression. Second, we used the EMR data from a single center. Validation using EMR data from various centers is required. Third, we utilized a machine learning model using EMR data at diagnosis and initial treatment. Therefore, this model can predict radiographic progression only when a patient first visits the hospital. In future, it will be necessary to develop a model that can predict radiographic progression at various time points by advancing machine learning models. Fourth, there may be models using algorithms that are better than the machine learning models developed in this study. It is possible to try a better model using an artificial neural network, but it may become more difficult for clinical application owing to the limitations of the “black box” model.

Among the datasets, including for the first, second, and third visits, predicting the radiographic progression of the second visit using the first visit dataset resulted in the best performance, with the highest accuracy and AUC. Therefore, the clinical features of the first visit are likely to contain important information for predicting radiographic progression. In terms of importance of features, mSASSS, age, ALP and CRP were ranked high. In addition to EMR data, various types of data, such as images and life-log, may be required to increase accuracy.

Ethics approval: The Hanyang University Seoul Hospital institutional review board approved this study (HYUH 2020-03-012-003). This study included only anonymized patient data and all research was performed in accordance with the Declaration of Helsinki.

Informed consent waiver statement: The Hanyang University Seoul Hospital institutional review board approved this study (HYUH 2020-03-012-003). Informed consent was waived because this study retrospectively reviewed the EMRs.

Availability of data and materials: The datasets used and/or analysed during the current study available from the corresponding author (Tae-Hwan Kim, [email protected]) on reasonable request. Transfer of data must be approved by the the Hanyang University Seoul Hospital institutional review board.

Competing interests: The authors declare that they have no competing interests

Funding: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT). (No. NRF-2021R1C1C1009815)

Author contributions: BSK, MJ, NK, and THK contributed to the study conception and design. All authors contributed to data acquisition, analysis, or interpretation. SL and KBJ scored the spinal radiographs independently. BSK, MJ, and NK were responsible for the statistical analyses. BSK and MJ drafted the manuscript and all coauthors were involved in critical revisions for maintenance of intellectual content. NK and THK provided administrative, technical, or material support. THK and NK had full access to all study data and takes responsibility for data integrity and data-analysis accuracy. All authors approved the final version to be submitted for publication.

Acknowledgments: We would like to thank all members of Biomedical Engineering in Asan Medical Institute of Convergence Science and Technology.

Inman RD: Axial Spondyloarthritis: Current Advances, Future Challenges. Journal of Rheumatic Diseases 2021, 28(2):55-59.
Brown MA, Li Z, Cao KL: Biomarker development for axial spondyloarthritis. Nat Rev Rheumatol 2020, 16(8):448-463.
Lorenzin M, Ometto F, Ortolan A, Felicetti M, Favero M, Doria A, Ramonda R: An update on serum biomarkers to assess axial spondyloarthritis and to guide treatment decision. Ther Adv Musculoskelet Dis 2020, 12:1759720X20934277.
Rademacher J, Tietz LM, Le L, Hartl A, Hermann KA, Sieper J, Mansmann U, Rudwaleit M, Poddubnyy D: Added value of biomarkers compared with clinical parameters for the prediction of radiographic spinal progression in axial spondyloarthritis. Rheumatology (Oxford) 2019, 58(9):1556-1564.
Fontanella S, Cucco A, Custovic A: Machine learning in asthma research: moving toward a more integrated approach. Expert Rev Respir Med 2021, 15(5):609-621.
Bressem KK, Vahldiek JL, Adams L, Niehues SM, Haibel H, Rodriguez VR, Torgutalp M, Protopopov M, Proft F, Rademacher J et al: Deep learning for detection of radiographic sacroiliitis: achieving expert-level performance. Arthritis Res Ther 2021, 23(1):106.
Deodhar A, Rozycki M, Garges C, Shukla O, Arndt T, Grabowsky T, Park Y: Use of machine learning techniques in the development and refinement of a predictive model for early diagnosis of ankylosing spondylitis. Clin Rheumatol 2020, 39(4):975-982.
Faleiros MC, Nogueira-Barbosa MH, Dalto VF, Junior JRF, Tenorio APM, Luppino-Assad R, Louzada-Junior P, Rangayyan RM, de Azevedo-Marques PM: Machine learning techniques for computer-aided classification of active inflammatory sacroiliitis in magnetic resonance imaging. Adv Rheumatol 2020, 60(1):25.
Joo YB, Baek IW, Park YJ, Park KS, Kim KJ: Machine learning-based prediction of radiographic progression in patients with axial spondyloarthritis. Clin Rheumatol 2020, 39(4):983-991.
Walsh JA, Pei S, Penmetsa G, Hansen JL, Cannon GW, Clegg DO, Sauer BC: Identification of Axial Spondyloarthritis Patients in a Large Dataset: The Development and Validation of Novel Methods. J Rheumatol 2020, 47(1):42-49.
Walsh JA, Pei S, Penmetsa GK, Leng J, Cannon GW, Clegg DO, Sauer BC: Cohort identification of axial spondyloarthritis in a large healthcare dataset: current and future methods. BMC Musculoskelet Disord 2018, 19(1):317.
Walsh JA, Rozycki M, Yi E, Park Y: Application of machine learning in the diagnosis of axial spondyloarthritis. Curr Opin Rheumatol 2019, 31(4):362-367.
Walsh JA, Shao Y, Leng J, He T, Teng CC, Redd D, Treitler Zeng Q, Burningham Z, Clegg DO, Sauer BC: Identifying Axial Spondyloarthritis in Electronic Medical Records of US Veterans. Arthritis Care Res (Hoboken) 2017, 69(9):1414-1420.
van der Linden S, Valkenburg HA, Cats A: Evaluation of diagnostic criteria for ankylosing spondylitis. A proposal for modification of the New York criteria. Arthritis Rheum 1984, 27(4):361-368.
Creemers MC, Franssen MJ, van't Hof MA, Gribnau FW, van de Putte LB, van Riel PL: Assessment of outcome in ankylosing spondylitis: an extended radiographic scoring system. Ann Rheum Dis 2005, 64(1):127-129.
Koo BS, Oh JS, Park SY, Shin JH, Ahn GY, Lee S, Joo KB, Kim TH: Tumour necrosis factor inhibitors slow radiographic progression in patients with ankylosing spondylitis: 18-year real-world evidence. Ann Rheum Dis 2020, 79(10):1327-1332.
Lee TH, Koo BS, Nam B, Oh JS, Park SY, Lee S, Joo KB, Kim TH: Conventional disease-modifying antirheumatic drugs therapy may not slow spinal radiographic progression in ankylosing spondylitis: results from an 18-year longitudinal dataset. Ther Adv Musculoskelet Dis 2020, 12:1759720X20975912.
Haroon N, Inman RD, Learch TJ, Weisman MH, Lee M, Rahbar MH, Ward MM, Reveille JD, Gensler LS: The impact of tumor necrosis factor alpha inhibitors on radiographic progression in ankylosing spondylitis. Arthritis Rheum 2013, 65(10):2645-2654.
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V et al: Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 2012, 12.
Breiman L: Random Forests. Machine Learning 2001, 45(1):5-32.
Sheridan RP, Wang M, Liaw A, Ma J, Gifford E: Correction to Extreme Gradient Boosting as a Method for Quantitative Structure–Activity Relationships. Journal of Chemical Information and Modeling 2020, 60(3):1910-1910.
Kim J-H: Estimating classification error rate: Repeated cross-validation, repeated hold-out and bootstrap. Computational Statistics & Data Analysis 2009, 53(11):3735-3745.
Greenwell BM, Boehmke BC, McCarthy AJ: A Simple and Effective Model-Based Variable Importance Measure. ArXiv 2018, abs/1805.04755.
Kuhn M, Johnson K: Applied Predictive Modeling; 2013.
Poddubnyy DA, Rudwaleit M, Listing J, Braun J, Sieper J: Comparison of a high sensitivity and standard C reactive protein measurement in patients with ankylosing spondylitis and non-radiographic axial spondyloarthritis. Ann Rheum Dis 2010, 69(7):1338-1341.
Ramiro S, van der Heijde D, van Tubergen A, Stolwijk C, Dougados M, van den Bosch F, Landewe R: Higher disease activity leads to more structural damage in the spine in ankylosing spondylitis: 12-year longitudinal data from the OASIS cohort. Ann Rheum Dis 2014, 73(8):1455-1461.
Koo BS, Lee S, Oh JS, Park SY, Ahn GY, Shin JH, Joo KB, Kim TH: Early control of C-reactive protein levels with non-biologics is associated with slow radiographic progression in radiographic axial spondyloarthritis. Int J Rheum Dis 2022, 25(3):311-316.
Castro-Zunti R, Park EH, Choi Y, Jin GY, Ko SB: Early detection of ankylosing spondylitis using texture features and statistical machine learning, and deep learning, with some patient age analysis. Comput Med Imaging Graph 2020, 82:101718.
Rajkomar A, Dean J, Kohane I: Machine Learning in Medicine. N Engl J Med 2019, 380(14):1347-1358.
Gao Y, Cui Y: Deep transfer learning for reducing health care disparities arising from biomedical data inequality. Nat Commun 2020, 11(1):5131.

No competing interests reported.

20220927supplFiguresBMC.docx

Download PDF

Journal Publication

published 20 Dec, 2023

Read the published version in Journal of Rheumatic Diseases →

Version 2

posted

You are reading this latest preprint version

Machine learning models with time-series clinical features to predict radiographic progression in patients with ankylosing spondylitis

Status:

Journal Publication

Version 2

Abstract

Background

Methods

Results

Conclusion

Figures

Background

Methods

Patients

Clinical data

Assessment of radiographic progression

Model design

Performance evaluation

Feature selection

Statistical analysis

Results

Differences between non-progressor and progressor groups

Predicting radiographic progression with three time-point datasets

Importance of features for predicting radiographic progression

Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 2